clm_30e5699670024d4ba64abc49238593c4_r1

2025-11-28_tpuv7-google-takes-a-swing-at-the::TPU7-0017

① SA Source

• Source: 開啟完整 SA 文章
• Section: Why use Google’s ICI 3D Torus Architecture?
• Line hint: 467

Context Before

Why use Google’s ICI 3D Torus Architecture?

But what is so great about Google’s unique ICI scale-up network – other than all the fancy cube diagrams one can spend countless hours drawing?

Evidence

World Size: The most obvious benefit is the very large 9,216 TPU maximum world size that the TPUv7 Ironwood supports.

Context After

Reconfigurable and Fungibility: The use of OCSs mean that the network topology inherently supports the reconfiguration of network connections to support a high number of different topologies – in theory thousands of topologies. Google’s documentation site lists out 10 different combinations (image earlier in this section), but these are only the most common 3D slice shapes – there are many more available.

Even slices of the same size can be reconfigured differently. In the simple example of a Twisted 2D Torus diagrammed below, we see how looping across to an index of a different X coordinate instead of an index of the same X coordinate can reduce the worst-case number of hops and the worst-case bisection bandwidth. This can help improve all to all collective throughput. A TPUv7 cluster will twist at the 4x4x4 cube level.

② Atomic Claim

TPUv7 Ironwood ICI 的核心優勢之一是最大 9,216 TPU scale-up world size。

• Epistemic Mode: ASSERTED
• Mapping Status: COMPLETE

③ Semantic Frame

{
  "attribute": "maximum_world_size",
  "context_nodes": [
    {
      "id": "04_knowledge_base/TPU",
      "label": "TPU"
    }
  ],
  "entity": {
    "id": "04_knowledge_base/TPUv7",
    "label": "TPUv7"
  },
  "frame_type": "ATTRIBUTE",
  "qualifiers": {
    "condition_text": null,
    "numeric_mentions": [
      "9,216"
    ],
    "temporal_mentions": []
  },
  "value": {
    "numeric_mentions": [
      "9,216"
    ],
    "value_text": "9,216 TPUs"
  }
}

④ Canonical Entity Mapping

Role	Surface Label	Canonical Target
entity	TPUv7	TPUv7
context_0	TPU	TPU

⑤ Human Review

請在 Properties 逐項確認：

• 原文 → Atomic Claim 是否忠實
• Atomic Claim → Semantic Frame 是否忠實
• Canonical Entity mapping 是否正確
• Epistemic mode 是否保留原文語氣
• 最後選擇 review_action

Review state

Markdown 內文不是正式 approval。只有 Apply bridge 寫入的 Decision Ledger event 才是正式決策。